Ludwig Wittgenstein believed that problems in philosophy are really problems in the interpretation of language. Nowhere today do we see this more clearly than in discussions about artificial general intelligence. Despite being one of the most important topics in the world, people’s definitions of “AGI” range from roughly last year’s coding agents all the way up to essentially God, and few people have tried to seriously disambiguate these. So I might as well give it a shot.
Level 0: Chatbots
Current frontier models are fundamentally no more capable than the chatbots of old. Depending on how strong a Level 0 someone is, they might draw this line at LaMDA, GPT-3, GPT-2, Amazon Alexa, Cleverbot, or ELIZA. LLMs are no more than stochastic parrots, blurry JPEGs of the Internet. They don’t actually think or reason, they just simulate it. Sure, the models are bigger, and they can store more facts and produce the illusion of intelligence better, but they aren’t even a new technology, let alone anything that would lead to sci-fi pipe dreams of utopia or doom1.
The industry that has popped up around AI, with its billion-dollar seed checks and trillion-dollar valuations and circular financing deals, is a giant scam that will run out of steam once people realize how useless the models are, and the bots will go away and we’ll all go back to normal.
Many Level 0-1s have (often exaggerated) critiques of AI unrelated to its perceived lack of capability: AI has a tendency to create slop that makes people dumber, it’s being controlled by fascist/woke Big Tech, it uses tons of energy and water, it’s trained on stolen data and exploits hidden labor, it enables scams/spam/deepfakes, etc.
Examples of Level 0 include Timnit Gebru, Ed Zitron, and Gary Marcus, though as real-world capabilities continue to advance, critics on this level are fewer and farther between.
This worldview would be roughly accurate if AI were at the capability level of 2019. It can be basically entirely dismissed today.
Level 1: Emerging Technology
LLMs are useful. They can be used to draft emails, write code, summarize long PDFs, and explain concepts to you. These are enough to spend $20 a month on a ChatGPT or Claude subscription and create real businesses built on the models’ APIs. Billions of dollars will be invested in the tech and the infra behind it, startups will blossom, and fortunes will be made. Even if we don’t get sci-fi AGI, our lives will get a little bit better and more convenient. But returns to scaling pre-training have diminished, we’re running out of data, and maybe LLMs can’t scale all the way to AGI.
On top of that, frontier labs’ valuations keep climbing into the hundreds of billions, big tech companies are spending so much on AI capex that they’ve created a global memory shortage, and while AI revenues are increasing, they aren’t increasing enough to make the labs profitable (at least not yet). The labs have committed so much to compute spend that if they don’t grow their revenues at an unprecedented rate, they won’t be able to afford to continue — the bubble will pop, and the market will crash. We might still get “true” AI eventually, but it’ll take a long time, with more AI winters on the way there.
You can compare the Level 1 view of AI to the current impact of technologies like crypto, VR, or quantum computing: many billions in investment subsidized by venture capital and big tech free cash flow, a lot of money made and lost, and many startups founded and killed off, but little to no real impact on the average person’s life.
This worldview would be roughly accurate if AI were at the capability level of 2023, but has been largely debunked by continued rapid revenue growth, demand for compute, and increases in fundamental model capabilities. The best example of a Level 1 is Daron Acemoglu, or perhaps a European VC.
Level 2: The Internet
LLMs are now very useful. They can complete major software engineering and white-collar labor tasks by themselves. Everyone can have their own personalized AI tutor, therapist, agent, companion, or boyfriend. The dream of an intelligent Siri or Alexa has finally become real. People pay for AI subscriptions just like they pay for Wi-Fi or Netflix, and enterprises will carve out a spending category for tokens alongside employee salaries. Eventually, compute buildouts and inference improvements will give us “intelligence too cheap to meter”, even if it’s not quite human-level intelligence. AI becomes a major industry, reaching a double-digit percentage of GDP and creating many trillions of dollars of market value.
Whether or not AI is a “bubble” in the moment doesn’t matter — if we see a market correction and some businesses go away, the technology is still real and will still compound over the coming decades. There are still major inefficiencies in both its implementation and the actual productivity boost provided to workers, and GDP growth will continue trudging along at 2% per year, maybe bumping up to 2.5 or 3%, helped by further diffusion and scaffolding improvements. AI might kill a few jobs — translator, copywriter — but on net, it will create a huge amount of jobs at higher wages than before, including in new categories that don’t exist yet. If political constraints allow it, we’ll finally get cheap energy, space exploration, and advances in healthcare. People will live to a hundred. The future is bright2. But it isn’t transformative.
This is the “AI as normal technology” view, and the view of much of the San Francisco tech ecosystem, especially “e/acc”3 and “builder” types. It would be roughly accurate if model capabilities plateaued around 2025-2026 levels.
Level 3: The Industrial Revolution.
Not even the most hardline skeptics can doubt AI’s capabilities anymore. It’s no longer “only” an incredibly useful tool that helps people do their jobs better, it’s now become a generalist agent that can do their jobs for them. First, advances in computer use and long-horizon task completion lead to a “drop-in remote worker”. Then, advances in robotics allow AI to substitute for human labor in the physical world as well. Organizational inertia, comparative advantage, and human preferences (e.g. for human babysitters, priests, or AI alignment researchers) keep many humans productively employed for a long time, and makes the rate of labor displacement go slower than the predictions of the most bullish people.
Regardless, with so many copies of roughly human-level AIs running much faster than humans and performing economically valuable tasks much cheaper, the economy starts to grow fast, much faster than the 2-3% a year we’re used to4. Any industry you can think of — law, medicine, software, engineering, design — soon turns into a dwindling group of human executives coordinating a vast and rapidly growing “population” of AIs. Science and research experience explosive growth in some areas and slower growth in others. We get our cheap energy, space exploration, and healthcare from Level 2, but along with it effective prevention and treatment of most physical and mental illness, incredibly cheap solar and nuclear power, major industry in space, infinitely customizable software, life extension, and so, so much more.
The rhythm of people’s daily lives will fundamentally change, too. You won’t have to go to work at your terrible retail or office job anymore — instead, you’ll earn capital income collected from the ever-growing AI economy5. There will be many vast, inconceivably complex structures on earth and in space housing AI compute, power generation, and factories. But life will also feel very normal in a lot of ways. There will be beautiful, wealthy, safe, clean cities for humans to live in. The Amalfi Coast will look just like how it did in 2026, or in 1500. You’ll spend lots of time with your family and friends, picking up new hobbies, appreciating incredible art, working hard on things that are meaningful to you. Eventually, it will feel no stranger than smartphones, cars, and industrial civilization feel to the descendants of hunter-gathering apes today.
The most rigorous view of Level 3 is Robin Hanson’s Age of Em6, or perhaps Dario Amodei’s Machines of Loving Grace.
This is the beginning of what the future could look like if AI capabilities7 improve over the coming years like they have over the previous years — the beginning of what it means to be AGI-pilled. Do you feel the AGI, anon?
Level 4: Homo sapiens.
To be AGI-pilled is not the same thing as to be ASI-pilled. A large population of roughly human-level AIs with superhuman speed and coordination would produce a crazy but still recognizable world. But what if the AIs are superhuman in cognition, too?
The economy, already seeing totally unprecedented growth at Level 3, becomes something more like a self-replicating machine, or a vast organism. There will still be fundamental abstractions between agents — trade, money, and so on — but coordination technologies will become so advanced that we’ll no longer be able to fully understand what’s going on8. Growth will start to move away from the limits of economics and more towards the limits of physics. Some bacteria can undergo mitosis in under 10 minutes. How quickly can self-replicating chip fabs or data centers do it?
Capabilities may well remain jagged. Maybe some things, like weather prediction and breaking encryption algorithms, are computationally intractable. Maybe you can’t reverse dementia if information has already been lost. But superhuman intelligence, working smarter and faster and in a more coordinated way, will be able to make huge progress on a staggering variety of problems. We’ll finally get true sci-fi tech, like nanotechnology, human intelligence enhancement, real full-dive virtual reality, total body revision.
Culture will stratify according to whether and how much different people decide to adopt these technologies, like the difference between the modern SF tech enthusiast, the median American, and the Amish, and the average person’s life changes accordingly. Some people will live Level 3 lives, recognizable to us today. Others will live in space or in virtual reality, or modify their bodies and minds into something more transhuman. Likely, no one will have to die anymore — extremely advanced biology will take care of that, and the dragon-tyrant will finally be no more.
Of course, a class of entities with this degree of power isn’t without its own issues. The biggest one is obsolescence — at this level, humans are totally superfluous to the economy, and the AIs could get away with killing us all if they wanted to. The key question, then, is how to make them not want to. This is the AI alignment problem.
When people read about futures like this, they often feel a degree of quiet unease, or future shock. But the people living in this future will likely be quite happy. If the AIs are aligned and things go well, it will be unbelievably awesome.
Level 4 and above gets very speculative. Few people have seriously engaged with this level of technology, but examples include Iain Banks’ Culture series and the epilogue of AI 2040.
Level 5: Life.
Now things really get different and weird. At this level, there are virtually no constraints on what technology can do, aside from physics itself. We might see all kinds of crazy stuff: black hole computers of unimaginable information density, swarms of nanobots capable of rewriting the surfaces of planets and the deepest values of brains, full-on posthumanism, countless simulated entities living on computational substrates, wild sci-fi all around9.
The economic, social, and cultural changes in such a world are essentially impossible to predict, as will be whatever layers of abstraction are above culture. The future will be as complex to us as we are to bacteria, for the most part utterly qualitatively beyond the comprehension of unmodified humans.
Most writing about Level 5 futures is less rigorous forecasting and more science fiction, with authors like Greg Egan, Olaf Stapledon, Stanisław Lem, and Charles Stross portraying some of it. For more of a nonfiction perspective, see Nick Bostrom or Nick Land.
The next time you hear the term “AGI” at an SF party or read it in an X thread, it’s critical to disambiguate what is actually being talked about. Is it a better chatbot? Is it a human-level coder? Is it a new species? Is it God?
Of course, AGI can progress from one level to another over time. Someone could believe in a Level 2 future for the next decade and Level 4 in 100 years. We are already between Level 1 and 2 in terms of impact, and we have so much farther to go. The models still can’t learn continually over time, they are still so much less sample-efficient than humans, they still have relatively poor taste in ideas compared to the best humans, and they’re still barely diffused throughout the economy.
This isn’t to say that a higher level of AGI-pilled-ness is automatically more correct. Maybe there really are bottlenecks on this stuff, and AGI will resolve at a Level 2 or 3. “Automated Walmart but not SpaceX or CERN.” But maybe, the trajectory of the next decade will look like the trajectory of the last one, and we’ll get a future unlike anything we’ve seen before.
Importantly, some people at lower levels (like Yann LeCun) believe that it’s just this paradigm that will cap out at Level 0-1, and future AI research directions will take us to Level 3+.
One aspect of Level 2 and upwards that’s harder to predict is the impact on culture. For example, the Internet fundamentally changed dating by mediating most interaction through dating apps, enabled the rise of remote work through videoconferencing, and shifted consumption habits away from books and TV towards first online streaming and then short-form video and social media feeds. What Level 2 AI would do here is somewhat speculative, but you could imagine:
better image and video models that can create more interesting and immersive entertainment experiences. Don’t like the ending of a TV show? Just change it. Or make your own from scratch, or have your AI make something you’d want.
more effective online dating thanks to AI matchmaking
social media communities (e.g. Discords) that are a mix of human users and AI agents (like the Cyborgism Discord server)
The original e/acc manifesto is much closer to Level 5-6, but most people with “e/acc” in bio on X are closer to a Level 2.
In Level 3, the regime of growth fundamentally changes, rather than just the rate of growth. The analogy here is how prior to the Industrial Revolution, the global economy grew around 0.1% a year, practically invisible growth even over a lifetime. In 1700s Britain, a remarkable series of achievements — cheap energy thanks to coal, mechanical production of goods, large and liquid capital markets, and pro-innovation institutions — caused the global economy to grow at a sustained rate of 2-3% per year, doubling every generation. This led to a rapid and unprecedented increase in living standards, life expectancy, population, energy capture, and individual liberty. It was the most important thing ever in human history. AI unlocks a new growth regime by allowing capital to substitute for labor, thus increasing the labor force arbitrarily with increased capital investment.
This is, of course, assuming that we solve the political and economic problems of automating much or all of human labor — in particular, that states are incentivized to provide welfare and avoid tyranny by the implicit potential of humans to withhold their labor or revolt entirely, and this constraint goes away in an AGI world. See The Intelligence Curse by Luke Drago and Rudolf Laine.
Age of Em, one of the most rigorous works of hard scenario forecasting ever written, examines a future where human brain emulation becomes possible, leading to a world of digital human copies, or “ems”. In our world, we’ll almost certainly get machine intelligence before human brain emulation, but LLMs are human enough that the book’s analysis still works. It doesn’t stay Level 3 forever, and explicitly includes the possibility of rapid economic growth and/or em intelligence enhancement leading to a Level 4 or 5 world. Regardless, the intermediate chapters are the best depiction I’ve seen of what a Level 3 world would look like.
Not AI diffusion, not AI adoption, not scaffolding — fundamental AI capabilities.
Some would say that this has already been the case for some centuries since the Industrial Revolution, with capitalism as a runaway self-accelerating process and the full economy far too complex for any human to understand, closer to an organism already. This idea was first articulated by Karl Marx and fleshed out in greater detail by Nick Land.
Some Level 5s even think that sufficiently powerful superintelligences could find exploits in physics, not unlike the fact that you can release a city-destroying amount of energy from a fission reaction, to do things like faster-than-light travel, but this is speculative.
Cool summation and look into the future!
This article doesn't actually disambiguate between common conceptions of AGI, instead it's disambiguating between forecasts of plateaued LLM utility?
These are two very different things?
For all their faults I don't think Gary Marcus or Timnit Gebru think AGI is definitionally "a giant useless scam", they just think LLMs are.